Guru Discover Change Context

---
name: guru-discover-change-context
description: Discover fresh current and archived change context, run the semantic evidence gate, and hand minimal current identity to Guru Team requirement clarification.
---

# Guru Discover Change Context

The current executing AI is this Skill's semantic owner. Read the complete
contract and perform its review yourself before authoring the current owner
result and calling record, check, and invoke. `owner_not_yet_executed` is an
internal state to continue this review, not a typed stop or a missing external
owner. Do not wait for another agent, agent ID, subagent evidence, or a
pre-existing owner result. Real missing authority or prerequisites still
follow this Skill's declared routes; runtime cannot supply your judgment.

Use this Skill after `guru-sync-base:synced`, or when a standalone caller asks
to discover change context from a fresh base with explicit issue, request,
path, command, config, schema, or symbol clues.

Load [references/contract.md](references/contract.md). Execute its semantic
closed loop in the declared order, complete the AI Review Gate before any
recorder/validator, then return exactly one declared typed exit.

Run commands from the reviewed repository root. The full execution package is
`.trellis/guru-team/skills/packages/guru-discover-change-context`; `scripts/`,
`schemas/`, and `examples/` references below resolve under that package, not
the `.agents/`, `.codex/`, `.claude/`, or `.cursor/` discovery projection.
Use its managed shell wrappers; do not import eval/private Python runtime to
construct recorder-owned fields. The contract lists the fields record derives.

Before current-state, duplicate, or history retrieval, read
`.trellis/spec/workflow/semantic-retrieval.md` and apply that shared contract in
this Skill's semantic evidence gate. A one-language zero result cannot establish
that a mechanism, duplicate, or historical decision is absent.

Use the dispatcher-only wrappers for history preview, owner-result recording,
and owner-result checking. Normal recording/checking is stdin/stdout-only and
does not create a repository artifact. Pass the active Discovery 2.0 input and
independent `base_current` transition to record, check, and invoke in one closed
semantic-owner 1.0 envelope via `--invocation -`. Include `schema_version`,
`public_input`, `transition`, `owner_context={}`, and `owner_result`; mode comes
from `public_input.mode`. Replace only `owner_result` with recorder stdout in
memory before check and invoke. No caller supplies or reconstructs a Sync
private result. See the exact Issue #384 migration in the contract: the former
three input locators and recorder `--mode` are removed, not alternate paths.
A caller-authored `refresh_base` result
records only the observed current stale codes, then reruns the complete Skill
through `guru-sync-base` and live authority.

For an active-task workflow invocation, pass `--active-task <task>` to record,
check, and public invoke so live validation binds `task.json.branch` and the
current task worktree without creating a checkpoint. Only when this same owner
is genuinely interrupted, also pass `--recovery-continuation-id <id>` to all
three calls. That additional identity lazily creates one minimal ignored
checkpoint, which a complete live-authority rerun checks and successful DTO
validation consumes.

Fail closed on stale base/live/blob/query/archive identity, invalid evidence,
unknown exits, or missing compatible runtime. This package is not
self-contained or portable.

After the semantic gate and owner recorder/checker complete, invoke
`scripts/invoke.sh --invocation -` with the closed call-local public input,
`base_current` transition, and checked owner result on stdin to serialize the
minimal handoff. The runtime reruns the existing checker and
derives the route from its checked `typed_exit`; callers never name the expected
exit. Consumers receive caller-owned continuation only, never the private
owner-result body.